Insertion of interjectory response based on prosodic information

This paper investigates 'aizuchi' responses (back-channel feedback, interjectory responses or oral nodding) to user's utterances by a computer. In telecommunications, 'aizuchi' responses are quite important to inform the hearer's condition to the speaker. In a man-machine telecommunication environment, user's comfortableness may improve if a dialog system can give suitable 'aizuchi' to user's utterances. For this purpose, it is important that a system detect conditions of 'aizuchi' insertion and insert one at appropriate timing. Firstly, occurances of 'aizuchi' were analyzed using a simulated dialog corpus of telephone-shopping-tasks. This showed that there are some prosodic cues in utterances in which 'aizuchi' is inserted. Then, timing of 'aizuchi' insertion was investigated and perceptual experiments were carried out. Lastly, we made a system which gives 'aizuchi' using prosodic information. This system detected suitable positions which agreed with original 'aizuchi's with 77% correct rate.